Add printing to check weights match torch version

.gitignore

@@ -129,3 +129,6 @@ poetry.toml
 
 # Scripts
 !/scripts/install-oneapi.bat
+
+# Test models for lora adapters
+/reduce-llms-for-testing

tests/test_lora_conversion_and_inference.sh

@@ -0,0 +1,127 @@
+#!/bin/bash
+set -e
+
+# Array of models to iterate over
+declare -a params=(
+    # "Gemma2ForCausalLM 64" TODO Add this model
+    "LlamaForCausalLM 64"
+    "Phi3ForCausalLM 64"
+)
+
+MODELS_REPO=reduce-llms-for-testing
+MODELS_REPO_URL=https://huggingface.co/ltoniazzi/$MODELS_REPO
+
+# Clone the Hugging Face repository if the directory does not exist
+if [ ! -d "$MODELS_REPO" ]; then
+    echo "Cloning the Hugging Face repository..."
+    git clone $MODELS_REPO_URL
+else
+    echo "Repository already exists. Skipping clone."
+fi
+
+# Load the expected starting strings from the text file
+EXPECTED_BASE=$(cat $MODELS_REPO/data/pale_blue_dot.txt)
+EXPECTED_LORA_HOT=$(cat $MODELS_REPO/data/bohemian_rhapsody.txt)
+EXPECTED_LORA_MERGED=$(cat $MODELS_REPO/data/bohemian_rhapsody.txt)
+
+# Declare a regular array to store results
+results=()
+
+trim_leading_whitespace() {
+    local input_string="$1"
+    # Remove leading whitespace
+    echo "${input_string#"${input_string%%[![:space:]]*}"}"
+}
+
+extract_starting_substring() {
+    local reference_string="$1"
+    local target_string="$2"
+
+    local target_length=${#target_string}
+    echo "${reference_string:0:$target_length}"
+}
+
+run_conversion_and_inference_lora() {
+    local model_name=$1
+    local size_matrix=$2
+
+    # Convert safetensors to gguf
+    echo "Running convert_hf_to_gguf.py for $model_name with size $size_matrix..."
+    python convert_hf_to_gguf.py $MODELS_REPO/$model_name/size=$size_matrix/base \
+    --outfile $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32.gguf \
+    --outtype f32
+
+    echo "Running convert_lora_to_gguf.py for $model_name with size $size_matrix..."
+    python3 convert_lora_to_gguf.py $MODELS_REPO/$model_name/size=$size_matrix/lora \
+    --base $MODELS_REPO/$model_name/size=$size_matrix/base \
+    --outtype f32
+
+    echo "Running llama-export-lora with lora for $model_name with size $size_matrix..."
+    ./llama-export-lora \
+    -m $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32.gguf \
+    -o $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32-lora-merged.gguf \
+    --lora $MODELS_REPO/$model_name/size=$size_matrix/lora/Lora-F32-LoRA.gguf
+
+    # Run inference
+    echo "Running llama-cli without lora for $model_name with size $size_matrix..."
+    OUTPUT_BASE=$(./llama-cli -m $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32.gguf \
+    -p "Look again at that dot." -n 50 --seed 42 --temp 0)
+
+    echo "Running llama-cli with lora for $model_name with size $size_matrix..."
+    OUTPUT_LORA_HOT=$(./llama-cli -m $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32.gguf \
+    --lora $MODELS_REPO/$model_name/size=$size_matrix/lora/Lora-F32-LoRA.gguf \
+    -p "I see a little silhouetto" -n 50 --seed 42 --temp 0)
+
+    echo "Running llama-cli with exported lora for $model_name with size $size_matrix..."
+    OUTPUT_LORA_MERGED=$(./llama-cli -m $MODELS_REPO/$model_name/size=$size_matrix/base/Base-F32-lora-merged.gguf \
+    -p "I see a little silhouetto" -n 50 --seed 42 --temp 0)
+
+    # Extract the corresponding substring from EXPECTED_BASE
+    # and remove initial white spaces in OUTPUT_BASE
+    OUTPUT_BASE=$(trim_leading_whitespace "$OUTPUT_BASE")
+    EXPECTED_BASE=$(extract_starting_substring "$EXPECTED_BASE" "$OUTPUT_BASE")
+    OUTPUT_LORA_HOT=$(trim_leading_whitespace "$OUTPUT_LORA_HOT")
+    EXPECTED_LORA_HOT=$(extract_starting_substring "$EXPECTED_LORA_HOT" "$OUTPUT_LORA_HOT")
+    OUTPUT_LORA_MERGED=$(trim_leading_whitespace "$OUTPUT_LORA_MERGED")
+    EXPECTED_LORA_MERGED=$(extract_starting_substring "$EXPECTED_LORA_MERGED" "$OUTPUT_LORA_MERGED")
+
+    # Compare the actual output with the expected start
+    if [[ "$OUTPUT_BASE" != "$EXPECTED_BASE" ]]; then
+        echo "Error: $model_name OUTPUT_BASE does not start with the expected string."
+        echo -e "Out=$OUTPUT_BASE\n\nExp=$EXPECTED_BASE"
+        exit 1
+    fi
+    if [[ "$OUTPUT_LORA_HOT" != "$EXPECTED_LORA_HOT" ]]; then
+        echo "Error: $model_name OUTPUT_LORA_HOT does not start with the expected string."
+        echo -e "Out=$OUTPUT_LORA_HOT\n\nExp=$EXPECTED_LORA_HOT"
+        exit 1
+    fi
+    if [[ "$OUTPUT_LORA_MERGED" != "$EXPECTED_LORA_MERGED" ]]; then
+        echo "Error: $model_name OUTPUT_LORA_MERGED does not start with the expected string."
+        echo -e "Out=$OUTPUT_LORA_MERGED\n\nExp=$EXPECTED_LORA_MERGED"
+        exit 1
+    fi
+
+
+    # Store the results in the regular array
+    results+=("
+    \n\033[1mResults for $model_name with size $size_matrix:\033[0m
+    \n\033[32m  • Base:\n$OUTPUT_BASE
+    \n\033[34m  • Lora hot:\n$OUTPUT_LORA_HOT
+    \n\033[36m  • Lora merged:\n$OUTPUT_LORA_MERGED
+    \n \033[0m
+    ")
+
+    echo "All steps completed for $model_name with size $size_matrix!"
+}
+
+# Run test for each model
+for param in "${params[@]}"; do
+    run_conversion_and_inference_lora $param
+done
+
+# Print all collected results
+echo -e "\n\n\033[1mSummary of All Results:\033[0m"
+for result in "${results[@]}"; do
+    echo -e "$result"
+done